Work
Applied ML
Receivables forecasting
~$950M in open invoices scored daily, median error 13 days
Ensemble of XGBoost and survival analysis predicting invoice payment timing, scoring 31K+ invoices across two years of daily batch runs. Dollar-weighted MAE of ~20 days against 95-day natural variability in collection times. Replaced a manual process that consumed hours of skilled financial analysts' time each cycle. Migrated the full train/score/monitor pipeline from Dataiku to containerized AWS: Docker images on ECR, scheduled ECS Fargate tasks, IAM scoping, Secrets Manager credential injection, CloudWatch logging with drift alerting. MLflow experiment tracking for model selection and rolling accuracy monitoring.
XGBoost · survival analysis · MLflow · Snowflake · Docker · ECR · ECS Fargate · IAM · CloudWatchInjury-risk model
155K+ SHAP explanations served to safety staff and site leaders
Weekly incident-probability predictions on active construction job sites with per-prediction SHAP factor decompositions stored alongside every score. Explainability-first design: the interface surfaces which features drove each site's risk so field safety teams can act on specific factors rather than an opaque number. Same containerized MLOps path (ECR/ECS/Fargate) as the forecasting system.
XGBoost · SHAP · ECS Fargate · SnowflakeEnterprise AI Platform
Data layer, security, and product discovery
40+ governed dynamic tables serving 1,300 employees and 761 projects
Owned the data layer for an enterprise AI platform, coordinating deeply with the client's data engineering team to integrate sources across a fractured landscape of ERPs, project management tools, HR systems, and manually maintained spreadsheets. Designed the OAuth security model (privilege-scoped tokens, role allow-listing, per-query connection lifecycle) that passed client InfoSec review. Led product discovery for client operation use cases that were later integrated into the platform.
Snowflake · Entra/Okta OAuth · product discoveryAI adoption module
Usage reporting across 1,300+ weekly active users, 550 custom GPTs
An $8B firm rolled out Claude, ChatGPT, Copilot, and Otter to 4,000 employees. Reporting was manual exports and hard-coded spreadsheets. I built the governed data layer that replaced it: coordinated cross-team ingestion into Snowflake, prototyped the pipelines, and made the product decisions on metrics and rollup structure for the dashboard.
Snowflake · dynamic tables · API ingestionProduct and Agent Work
Agent plugin marketplace
Two marketplaces (52 internal skills, 5 client-facing), phased rollout across two companies
Designed and built the agent skills program end-to-end: two versioned marketplaces (one for the internal team, one executive-focused for the client's 4,000-person org), a release pipeline with CI, and a QA process gating each skill on real-data validation before client deployment. Authored domain intelligence skills spanning finance, safety, sales, precon, recruiting, quality, and operations. Built the daily-brief family (morning intelligence digests) and subsidiary-specific variants (competitor intel, deal briefs). Designed the phased rollout strategy: internal adoption first with NPS gates, then client rollout coordinated with existing training programs. Ported skills across platforms (Claude CoWork, ChatGPT) to maximize reach across the org.
Claude · ChatGPT · Snowflake semantic views · CI/CD · GitHubExecutive agent deployments
Personal AI agents for the client's founder/chairman and CEO
Deployed and administered dedicated agent instances on AWS: ECR image management, Cloudflare tunnel routing, VNC-gated access, instance hardening. Ongoing administration including version upgrades and security patches.
AWS ECR · Docker · Cloudflare · security hardeningProduct discovery
Multiple use cases from field feedback to platform roadmap
Led product discovery for construction operations use cases, balancing off-the-shelf AI tools (ChatGPT Actions, custom GPTs) with custom-built backends to optimize time-to-value. Use cases included scope-of-work drafting and first-pass review using company standards as a knowledge base, architecture cost engineering comparison (tracking price drivers between estimate stages), and a preconstruction subcontractor module (validated data coverage across 1,100+ projects, coordinated with the data engineering team on the subcontractor data layer). Built RAG backends where off-the-shelf tools fell short, iterated from live field pilot feedback with PMs and project directors.
FastAPI · RAG · OpenAI vector search · GPT Actions · Snowflake